ShuaKang/robotwin_pill_rara

RoboTwin 2.0 move_pillbottle_pad: target demonstrations + retrieved pre-training subset (RARA, one task)

Everything needed to adapt a pre-trained Diffusion Policy to one RoboTwin 2.0 target task with FT, Co-FT and RARA, packaged for reproduction on a single task.

path	content
dp_zarr/narrow_move_pillbottle_pad.zarr	50 target demonstrations (7,324 frames), demonstration configuration (object position range 0.5 of the task window)
dp_zarr/ret250L_move_pillbottle_pad_ws2k.zarr	250 retrieved pre-training episodes (38,588 frames), all place_a2b_right
dp_zarr/ret250L_move_pillbottle_pad_ws2k.episodes.json	for each retrieved episode: index in the 2000-episode pool, source task, DTW distance, rank
geometry/*.npz	per-frame 5-d grasp-relative descriptor (desc, lens) for the target and the retrieved subset, used by RARA Stage 1

Zarr format (Diffusion Policy replay buffer): data/head_camera, data/right_camera (240x320x3 uint8), data/state, data/action (14-d joint position, action[t] = state[t+1]), meta/episode_ends. Conventions: RoboTwin 2.0 Aloha-AgileX, right arm only, head camera D435 at the "D4" pose + right wrist camera, robot body and left arm hidden; scene randomisation with 10 backgrounds and 2 clutter objects.

How the 250 episodes were selected

The pre-training pool is dp_zarr_ws/pre_ws_2k.zarr in ShuaKang/my_roboTwin2.0_training: 2000 demonstrations (place_a2b_right 1000; click_bell, open_laptop, rotate_qrcode, shake_bottle 250 each), every object sampled in one right-arm workspace x [0.05, 0.31] m, y [-0.20, 0.10] m, which covers this task's ID and OOD object ranges.

1. Task / language filter: keep only the pool episodes of the target's skill (pick-and-place -> place_a2b_right, 1000 episodes).
2. Rank them by the mean DTW distance between their grasp-relative descriptor trajectory (resampled to 32 points, standardised with pool statistics) and the 50 target demonstrations; keep the 250 nearest.

Use

• Pre-trained policy P: Diffusion Policy trained on the full 2000-episode pool (60 epochs, batch 256).
• FT: fine-tune P on the 50 target demos (300 epochs).
• Co-FT: fine-tune P on target + the 250 retrieved episodes (64 + 64 samples per batch).
• RARA: Stage 1 aligns the vision encoders on target frames against geometry-matched frames of the retrieved data (contrastive, 4 nearest positives), Stage 2 fine-tunes on the target with a feature anchor to P on the retrieved frames. In the reference runs Stage 1 drew its positives from the whole 2000-episode pool (pre_ws_2k.npz); using only this 250-episode subset is the self-contained variant.
• Evaluation: ID = demonstration configuration (cov_narrow), OOD = object range 0.65 of the task window (cov_ood_wa065); last three checkpoints x two seed batches x 25 episodes = 150 rollouts per split.

Code: https://github.com/Shua-Kang/my_roboTwin2.0 (environment and configs) and the robotwin_dp release (training and evaluation).